A Prediction of Antibiotic Resistance with Regard to Urinary and Respiratory Tract Infections

(1)

A

Prediction

of

Antibiotic

Resistance

with

Regard

to

Urinary

and

Respiratory

Tract

Infections

Felicia

Wallnäs,

Bella

Sinclair,

Stella

Belin,

Erik

Olby,

Hampus

Söderberg

Beställare:

Q-linea

Beställarrepresentant:

Fredrik

Pettersson

Handledare:

Lena

Henriksson

(2)

Abstract

In this project we set out to find when the resistance level against first line antibiotics would reach 20%. This was executed by first defining relevant bacteria and antibiotics for urinary and respiratory tract infec-tions (UTI’s, RTI’s). The data was collected from the European Center for Disease Control (ECDC) and the Center for Disease Dynamics, Economics & Policy (CDDEP). The data included the level of resistance for specific years for countries in Europe, as well as for the USA. A prediction model was made using the programming language R. A linear model was used to make a five and ten year prediction. The accuracy was tested. The results were then visualized using R and MATLAB.

The results show a big variation between different bacteria and antibiotic combinations. For the two Es-cherichia coli combinations the resistance is already near 20% for many countries and the resistance is increasing. For the three Klebsiella pneumoniae combinations the resistance is high in Southern Europe, meaning many countries have reached or are near 20%. For the two Pseudomonas aeruginosa combinations there is also a higher resistance in Southern Europe but the resistance is decreasing in most countries. The resistance for Enterococcus faecalis is also decreasing and is generally very low in all of Europe. For the only RTI relevant combination, Streptococcus pneumoniae and penicillins, the resistance is low and many countries except for Sweden show a decrease in resistance.

The USA did not have data for the same time span as Europe and was therefore analyzed separately. For many combinations the USA are near the 20% limit. Only for two combinations the USA showed a decrease in resistance level, and for one of those combinations the prediction is too uncertain to make any assumptions about. For the USA there were two more combinations for RTI than for Europe. For the S. pneumoniae and penicillins combination they have, just as most of Europe, a decreasing resistance. The two combinations with Acinetobacter spp. have a high resistance that is increasing.

The main challenge during this project was finding relevant data with a long timespan and with high cer-tainty. The data found is based on invasive isolates which means that the disease which the samples are taken from is not known. The timespan and the certainty of the data affected the accuracy of the prediction model and how long period that could be predicted. The prediction model generated 202 predictions that were visualized.

An ethical analysis was made concerning both research ethics and general ethics on the topic of antibiotic resistance. This analysis is meant to acknowledge these questions since we believe they are important when discussing antibiotic resistance.

The objective of this project turned out to be more difficult to attain than first believed. This was because of the lack of quality data. Even though we cannot give a clear answer when each country will reach a resistance of 20% this report gives a good understanding of how the situation looks for UTI and RTI relevant bacteria.

(3)

1 Acknowledgements . . . 4

2 Abbreviations . . . 4

3 Introduction . . . 5

3.1 Purpose and Motivation of the Project . . . 5

3.2 Antibiotic Resistance and Consumption . . . 5

3.3 Types of Samples . . . 6

3.4 Predictive Model . . . 6

4 Data Collection . . . 7

5 Results and Analysis . . . 8

5.1 Summary of Results . . . 8

5.2 UTI Relevant Combinations in Europe . . . 8

5.2.1 Enterococcus faecalis and Aminopenicillins . . . 9

5.2.2 Escherichia coli and Fluoroquinolones . . . 11

5.2.3 Escherichia coli and Third Generation Cephalosporins . . . 13

5.2.4 Klebsiella pneumoniae and Aminoglycosides . . . 15

5.2.5 Klebsiella pneumoniae and Fluoroquinolones . . . 17

5.2.6 Klebsiella pneumoniae and Third Generation Cephalosporins . . . 19

5.3 UTI and RTI Relevant Combinations in Europe . . . 21

5.3.1 Pseudomonas aeruginosa and Aminoglycosides . . . 21

5.3.2 Pseudomonas aeruginosa and Fluoroquinolones . . . 23

5.4 RTI Relevant Combinations for Europe . . . 25

5.4.1 Streptococcus pneumoniae and Penicillins . . . 25

5.4.2 Acinetobacter spp. and Aminoglycosides or Fluoroquinolones . . . 26

5.5 UTI Relevant Combinations for the USA . . . 26

5.5.1 Enterococcus faecalis and Aminopenicillins . . . 27

5.5.2 Escherichia coli and Fluoroquinolones . . . 27

5.5.3 Escherichia coli and Third Generation Cephalosporins . . . 27

5.5.4 Klebsiella pneumoniae and Aminoglycosides . . . 27

5.5.5 Klebsiella pneumoniae and Fluoroquinolones . . . 27

5.5.6 Klebsiella pneumoniae and Third Generation Cephalosporins . . . 27

5.6 UTI and RTI Relevant Combinations in the USA . . . 27

5.6.1 Pseudomonas aeruginosa and Aminoglycosides . . . 27

5.6.2 Pseudomonas aeruginosa and Fluoroquinolones . . . 28

5.7 RTI Relevant Combinations for the USA . . . 28

5.7.1 Streptococcus pneumoniae and Penicillins . . . 28

5.7.2 Acinetobacter spp. and Aminoglycosides . . . 28

5.7.3 Acinetobacter spp. and Fluoroquinolones . . . 28

5.8 A More Detailed Analysis of Four Countries . . . 28

5.9 Fitness of Model . . . 30

(4)

6.1 Data Collection . . . 32

6.1.1 Availability of Data . . . 32

6.1.2 Invasive Isolates . . . 32

6.2 Predictive Model . . . 32

6.2.1 Possible Improvements . . . 32

6.2.2 Alice’s Croquet Theory . . . 33

6.2.3 Reach of Our Prediction . . . 33

6.3 Presenting Our Data . . . 34

7 Conclusion . . . 34 8 Ethical Analysis . . . 35 8.1 Communicating Research . . . 35 8.1.1 Theoretical Aspects . . . 35 8.1.2 Application of Theory . . . 35 8.2 Responsibility . . . 36

8.2.1 The Researcher’s Responsibility . . . 36

8.2.2 The Politician’s Responsibility . . . 36

8.2.3 The Individual’s Responsibility . . . 37

8.2.4 The Doctor’s Responsibility . . . 37

8.3 Aspects of Risk and Rights . . . 37

8.3.1 The Individual versus Society . . . 37

8.3.2 Antibiotics as a Right . . . 38 8.4 Final Thoughts . . . 38 9 Methodology . . . 39 9.1 Database . . . 39 9.2 Predictive Model . . . 39 9.3 Visualization . . . 39 10 List of References . . . 41

(5)

1 Acknowledgements

Firstly we would like to express our appreciation to Lena Henriksson, Student faculty coordinator at the Biology Education Centre (IBG) at Uppsala University, our supervisor for her significant guidance during this project. She also ensured that we got several relevant lectures that further aided the project.

We also want to thank Tobias Jakobsson, Project manager at the Biology Education Centre at Uppsala Uni-versity, for giving us a lot of help with setting up our database on the Biology Education Centre’s servers. He has been constantly easy to contact and willing to thoroughly explain what we needed to do in order to use the database.

Furthermore we would like to thank Jan Kudlicka, PhD student at the Department of Information Technol-ogy at Uppsala University, for also helping us with our database. He provided us with aid in the design of the database, ensuring that we would store our data in an effective way.

Our appreciation is also extended to Rolf Larsson, Professor at the Department of Mathematics at Uppsala University, who helped us with our predictive model. He took the time to sit down and discuss our best approach and helped grasp some of the concepts needed to carry out this project.

We also want to extend our gratitude to Fredrik Pettersson, Researcher at the Department of Cell and Molecular Biology at Uppsala University, who has been our way of communication with our client Q-linea. In addition he has also given guidance in the direction of the project.

In addition we would like to thank our client, Q-linea. Without them there would not have been a project at all. They have been a very helpful client, always answering questions and providing feedback when asked. Finally we want to extend our gratitude to our peer-reviewers. Their feedback and input has greatly helped us to improve our work both in how it should be carried out as well as how it should be presented.

2 Abbreviations

CDC - Centers for Disease Control and Prevention CDDEP - Center for Disease Dynamics, Economics & Policy ECDC - European Centre for Disease Prevention and Control EFSA - European Food Safety Authority

EMA - European Medicines Agency

IBG - Biology Education Centre (Institutionen för biologisk grundutbildning) RSN - ResistanceMap Surveillance Network

RTI - Respiratory Tract Infections SQL - Structured Query Language TSN - The Surveillance Network UTI - Urinary Tract Infections WHO - World Health Organization WMA - World Medical Associations

(6)

3 Introduction

3.1 Purpose and Motivation of the Project

The purpose of this paper is to provide assistance to our client, Q-linea, to make an informed decision regard-ing in what direction they should develop their products. Q-linea develops products related to diagnosregard-ing bacterial infections as well as their resistance to antibiotics. For this project they are particularly interested in urinary tract infections (UTI) and respiratory tract infections (RTI). So to this end, they want to know if and when the relevant combinations of bacteria and antibiotics reach a resistance level of 20% in different parts of the world.

Q-linea is a company in Uppsala who are developing a product called ASTar (Antimicrobial Susceptibility Testing), which is a machine that enables faster diagnosis of sepsis. The diagnostic tools used today need a couple of days to identify the antimicrobial susceptibility in a patient who has sepsis. With ASTar, Q-linea is looking to decrease this time to a few hours. This will make it easier to decide which antibiotics should be used to treat a patient, and therefore also create the opportunity to reduce the overconsumption of antibiotics.

3.2 Antibiotic Resistance and Consumption

Antibiotic resistance is a global issue that right now is growing. One of the contributing factors to this, which is the most relevant for this paper, is the difficulty of identifying the pathogen which is causative of a disease. This sometimes forces doctors to empirically try to choose the correct antibiotics based on symptoms, or wait for a few days for a diagnosis. Since this can cause improper antibiotics to be prescribed, another one has to be used and so on until the pathogen is treated correctly. This is not optimal and contributes to the overconsumption of antibiotics.

Antibiotic usage does not only include the human usage but also usage in animals. An integrated analysis (EFSA 2015) by European Centre for Disease Prevention and Control (ECDC), European Food Safety Au-thority (EFSA), and European Medicines Agency (EMA) found that:

“In both humans and animals, positive associations between consumption of antimicrobials and the corresponding resistance in bacteria were observed for most of the combinations investigated. In some cases, a positive association was also found between antimicrobial consumption in ani-mals and resistance in bacteria from humans.”

Seemingly, both antibiotic consumption in animals and in humans are relevant to antibiotic resistance. Furthermore, a paper by Boeckel et al. (2015) predicts that the antibiotic consumption will be 67% higher by 2030 due to a “growth in consumer demand for livestock products in middle-income countries and a shift to large-scale farms where antimicrobials are used routinely”. Antibiotic resistance spreads between borders, due to food export and travel. Therefore a change in one country could have a global effect. This means that there are several economical and political factors that will affect the development of antibiotic resistance. These are near impossible to predict.

(7)

3.3 Types of Samples

For this paper the majority of the data collected are from invasive isolates. Invasive isolates can be defined as a sample acquired from a normally sterile body site (Crespo-Ortiz et al. 2014). For the data relevant to this paper this is further limited to isolates from blood and cerebrospinal fluid (European Centre for Disease Prevention and Control 2017). Cerebrospinal fluid is a fluid that can be found in and around the hollow spaces of the spinal cord and brain, and also between between two of the protective tissue layers of the brain (NCI Dictionary of Cancer Terms 2018).

Interpreting data from invasive isolates has both advantages and disadvantages that need to be taken into consideration. Invasive isolates makes the data more consistent since it avoids the problems of different clinical definitions, different sampling frames, or heterogeneous health care utilization. But they are not necessarily a good representation of the same bacteria in another infection. In the case of this paper this means that the resistance data may not be representative of a UTI or RTI, even if the bacteria and antibiotics are (European Centre for Disease Prevention and Control 2017). Even though different strains of the same bacteria can cause different infections and also have different resistance patterns, the data was divided by bacteria type and not by strain.

3.4 Predictive Model

A predictive model is a model that is fitted to previous data and then used to predict future data points. When predicting the future the prediction is called a forecast. The type of model used in this project were a linear model which resulted in a point forecast with a confidence interval. The point forecast is the line in the middle of the confidence interval. The time span the model predicts is called a prediction interval.

(8)

4 Data Collection

Since the purpose of the project is to look at antibiotic resistance with a focus on UTI’s and RTI’s, the first thing that needed to be defined was which bacteria most often cause these infections as well as which antibiotics are most likely prescribed for those situations. See Table 1 for the relevant bacteria for UTI’s (Imam 2016) and RTI’s (Sethi 2017) and their respective antibiotic. The antibiotics examined in this report were largely based on which antibiotics there were data on (European Centre for Disease Prevention and Control n.d.).

Table 1: Bacteria and antibiotics judged to be relevant for this project.

Bacteria Disease Antibiotic

Enterococcus faecalis UTI Aminopenicillins

Escherichia coli UTI Fluoroquinolones

Escherichia coli UTI Third generation cephalosporins Klebsiella pneumoniae UTI Aminoglycosides

Klebsiella pneumoniae UTI Fluoroquinolones

Klebsiella pneumoniae UTI Third generation cephalosporins Pseudomonas aeruginosa UTI/RTI Aminoglycosides

Pseudomonas aeruginosa UTI/RTI Fluoroquinolones Streptococcus pneumoniae RTI Penicillins

Acinetobacter spp. RTI Aminoglycosides

Acinetobacter spp. RTI Fluoroquinolones

The data needed for our prediction model for each combination of antibiotic and bacteria was resistance level, year, country, the number of tested isolates (N-value) or the number of tested positive isolates (R-value). The data should preferably, be presented in tabular formats (i.e. .xlsx, .csv) to automatically transfer the data into our database, thereby, minimizing the error.

The data collected from Europe was found at the ECDC webpage, which is an agency coordinated by the European Union to which many European countries report data. Because ECDC contained comprehensive data collected for a number of years, all data regarding Europe was collected from the ECDC. The data was available in .xlsx format and downloaded.

The data collected from the USA was obtained from the Surveillance Network (TSN), ResistanceMap Surveil-lance Network (RSN) (ResistanceMap 2018), and the Center for Disease Control National Healthcare Safety Network. RSN presented resistance levels in the form of maps, charts and trends. The values from the trends were transferred manually into a .xlsx format.

For each country and for each combination of bacterium and antibioticum, the number of tested isolates and the number of years containing data were evaluated. This was done in order to increase the quality of our result. We set the minimum limit for the number of years containing data to ten years. The limit of ten years was considered arbitrary due to our available data. Of those ten years the number of tested isolates needed to be 50 or higher for each year. The limit of number of tested isolates was based on the limit ECDC chose for their earlier analysis of antibiotic resistance (European Centre for Disease Prevention and Control n.d.). If combinations of bacteria and antibiotics for a country did not fulfill these limits, they were excluded.

(9)

5 Results and Analysis

In this part a general picture is given of which and how many countries have already reached 20%, will reach 20% and those that will not reach 20% in the prediction interval. Some combinations of bacteria and antibiotics have an uncertain prediction or a decreasing trend, which will also be presented. With uncertain we mean that the confidence interval is large. The most common cause of why the prediction was uncertain was that there was a high variance in the data, meaning that the resistance varied a lot from year to year and is not following a clear trend.

A five year prediction was made and was plotted in graphs. We also made a ten year prediction to show what would happen if the resistance follows our prediction model.

5.1 Summary of Results

It is difficult to say what year the resistance will reach 20% for the different combinations, since many coun-tries have already reached 20% and few pass the limit in our prediction interval. What we can say is that for some combinations relevant for UTI’s the resistance in Europe is low and the resistance is even decreasing. This is the case for the combination E. faecalis and aminopenicillins. For P. aeruginosa and fluoroquinolones some countries are above 20% but will probably pass below 20%.

For E. coli and third generation cephalosporins we see a steady increase. Even though many countries are below 20%, this combination might not be a good choice of treatment in the future. For E. coli and fluo-roquinolones the future does not look bright. Most countries are already above 20% and others are following. When looking at RTI’s there is only one combination that has good enough data and that is S. pneumoniae and penicillins. The resistance is very low and most countries have a decreasing resistance. If the trend is decreasing the predicted values might be negative. This is not applicable to real life since antibiotic resistance cannot be negative. The model does not take this into consideration.

In the USA the latest data is much older than for Europe resulting in a prediction only to 2017 or 2019. Overall the USA follows the same trends as Europe. Only the E. coli and fluoroquinolones combination is clearly above 20% and the other combinations are either below or around 20%. The USA also have a high number of isolates tested each year. It has data for Acinetobacter spp. from 1999 unlike Europe but the resistance is high for both types of antibiotics, significantly above 20%.

The conclusion we can draw is that there is not enough resistance data in neither Europe nor the USA in order to make a prediction that attains the project objectives. To get a basis for which antibiotics should be used when recommending antibiotics, a new approach should be used. In ten years we might be able to make a prediction which can attain the objective, but right now the quality data needed for that is missing.

5.2 UTI Relevant Combinations in Europe

The data in Europe began to be collected around 2000 or 2005 depending on the combination of antibiotics and bacteria. All countries have data from that point until 2016. That means that 2017 and 2018 are predicted since there is no data for those years yet. The five year prediction extends to 2021 and the ten year prediction to 2026.

(10)

5.2.1 Enterococcus faecalis and Aminopenicillins

Figure 1: Maps over Europe for 2010, 2015, 2020 and 2025 showing the resistance for E. faecalis against aminopenicillins in different countries. The resistance is generally low for this combination. Note that Italy has a sharp decrease between 2010-2015 but is predicted to increase after that. This is because Italy has very varying data going from high values to low values, making the prediction uncertain.

Most countries have data from 2001 and have a stagnating or a decreasing resistance trend. Generally the resistance is very low and only a few are above 20% including the UK. 19 countries will not pass 20% and for the UK it is too uncertain to say. In Table 2 all countries that are below 20% are presented with their resistance for 2016 and their predicted resistance in 2021.

In Figure 1 the maps for 2010, 2015, 2020 and 2025 are found. Almost all countries have a very low resistance and only the UK has an uncertain prediction that is above 20%. Portugal and Italy are slowly decreasing. In 2006, Bulgaria was above 20% but since then it has been below 20%. The prediction is uncertain but as of now it seems like Bulgaria is following a decreasing trend.

The resistance in the UK has been around 5% from 2005 to 2014, and then in 2015 it was over 35%. This makes the prediction very uncertain. For 2016 the resistance is at about 15% so it might pass 20% in the

(11)

coming years.

Table 2: The results from the prediction for E. faecalis and aminopenicillins. This table only contains the countries that have not yet reached 20% because there are no countries above 20%. For all countries the resistance level in 2016 and the predicted resistance in 2021 is shown.

Countries below 20% Resistance in 2016 (%) Resistance in 2021 (%)

Austria 0.3 0.1 Belgium 0.4 1.3 Bulgaria 6.1 -2.7 Croatia 7.8 7.1 Czech Republic 1.2 -0.3 Denmark 0.7 0.7 Finland 0.5 -0.5 France 0.3 -0.9 Germany 0.3 -0.3 Greece 3.7 2.4 Hungary 1.1 0.7 Ireland 0.3 -0.7 Italy 2.1 7.7 Netherlands 0.6 -0.04 Norway 0.2 -0.1 Portugal 0.9 10.5 Slovenia 0.6 1.2 Spain 0.4 1.3 Sweden 0.5 0.2

(12)

5.2.2 Escherichia coli and Fluoroquinolones

Figure 2: Maps of Europe for 2010, 2015, 2020 and 2025 showing the resistance for E.coli against fluoroquinolones in different countries. The resistance is high in all of Europe.

Generally, most countries have data from 2001 and are increasing. Some have passed the 20% limit long ago and some are getting closer. Seven countries including all of the Nordic countries will not pass 20% in the prediction. 19 countries have already passed 20%. For Germany, France, Lithuania and the UK it is too uncertain to draw any conclusions. In Table 3 the countries that have and have not passed 20% are shown. In Figure 2 the resistance maps for this combination are shown. The resistance was already high in 2010 and in 2020 we predict that all but Finland, Iceland, Norway and Sweden will have a resistance level above 16%.

Estonia is steadily rising and will probably reach 20% after 2021 but it is not seen in the prediction. The Netherlands is getting closer to 20% and might pass 20% in the coming years but the point forecast does not pass 20% in the prediction.

France had a steady increase from 2002 to 2009 and has since stagnated just below 20%, but the uncertainty is too big to say anything. Germany has been above 20% for some years but in 2015 and 2016 the country

(13)

dropped below. This makes the prediction uncertain. Lithuania passed 20% in 2015 and in 2016 dropped below again. The model predicts that it already has passed 20% or will pass the limit soon but the variance in the data makes the prediction uncertain. Therefore, we cannot say anything about when Lithuania will pass 20%. The UK has stagnated just below 20% since 2006 making the prediction uncertain. Therefore, we cannot say when or if the UK will pass 20% in the next few years.

Ireland has since 2006 stagnated just above 20% and will probably stay above but there is no trend as of right now. The data for Latvia is very scattered but in 2015 it passed the 20% limit.

Table 3: The results from the prediction for E. coli and fluoroquinolones. The first part of the table contains the countries that have not yet reached 20% and the second part the countries that have. For all countries the resistance level in 2016 and the predicted resistance in 2021 is shown.

Denmark 11.0 16.7 Estonia 13.9 18.1 Finland 11.5 14.9 Iceland 9.6 14.1 Netherlands 12.6 19.2 Norway 10.9 16.0 Sweden 13.7 16.3

Countries above 20% Resistance in 2016 (%) Resistance in 2021 (%)

Austria 19.8 26.8 Belgium 24.5 32.1 Bulgaria 42.2 50.7 Croatia 27.9 32.2 Cyprus 47.0 56.8 Czech Republic 27.6 31.1 Greece 32.1 43.4 Hungary 26.8 41.5 Ireland 22.9 33.0 Italy 43.9 55.6 Latvia 27.8 30.4 Luxembourg 28.9 36.7 Malta 41.5 43.2 Poland 33.1 41.9 Portugal 28.9 34.4 Romania 30.6 42.2 Slovakia 40.4 58.3 Slovenia 25.6 30.7 Spain 32.8 42.1

(14)

5.2.3 Escherichia coli and Third Generation Cephalosporins

Figure 3: Maps of Europe for 2010, 2015, 2020 and 2025 showing the resistance for E. coli against third generation cephalosporins in different countries. The resistance increases in all countries and many reach 20% resistance.

We would say that this is the combination with the best prediction because it has narrow confidence interval for most countries. The data started to be collected around year 2000 and the last record is from 2016. Most countries have a clear increase with a somewhat certain prediction, however some predictions are too uncertain to draw any conclusions from. 16 countries have not reached 20% at the end of the prediction. Seven countries might pass 20% in the coming years and five countries have already passed 20%. For Latvia, Lithuania, and Malta the prediction is too uncertain to say if they are above or below 20%. In Table 4 the countries that have not reached 20% and those that have can be seen. In Table 5 the countries that we predict will pass 20% in the next few years are listed.

In Figure 3, four maps are shown for 2010, 2015, 2020 and 2025. These are based on the ten year prediction. What we can see is that Southern and Eastern Europe has a higher resistance than Northern Europe, and with time the resistance will rise for all countries. Not all countries reach 20% in this prediction but the future does not look promising.

(15)

The prediction for Poland is uncertain but even with a large uncertainty the resistance will probably not reach 20% until after 2021. For Sweden the slope is a bit exponential which the linear model does not capture. The model does not predict that Sweden will reach 20%, but if the slope is truly exponential the prediction might not be true.

The 95% confidence interval for Croatia exceeds 20% in 2019 but the point forecast lies below, meaning that there is a possibility that Croatia reaches 20% in the next few years. Czech Republic will according to our prediction reach 20% somewhere between 2017 and 2021, the point forecast reaches 20% in 2020. Greece passed the 20% limit in 2013 but dropped below the line again. According to our prediction, Greece probably reached 20% again as of 2017. Hungary has according to the prediction, just as Greece, probably already reached 20% in 2017. The point forecast for Portugal passes 20% in 2021 but the resistance can reach 20% any time between 2017-2021. Spain might reach 20% between 2018 and 2021 but the point forecast does not exceed 20% in the prediction interval.

Latvia reached 20% in 2016 and the prediction is very uncertain, meaning that you cannot make any assumptions even though the point forecast passes 20% in 2018. Lithuania has the same problem with an uncertain prediction. Malta passed 20% in 2008 but then dropped to between 10-15%. The prediction is therefore very uncertain and no conclusion can be drawn. The uncertainty for Romania is very large but despite this we can say that the resistance have already reached 20% since Romania has been above 20% since 2011.

Table 4: The results from the prediction for E. coli and third generation cephalosporins. The first part of the table contains the countries that have not yet reached 20% and the second part the countries that have. For all countries the resistance level in 2016 and the predicted resistance in 2021 is shown. Some countries like Spain are close to the 20% limit in the last prediction year but does not pass 20% in the prediction interval.

Austria 10.0 14.7 Belgium 10.5 12.5 Croatia 14.7 16.4 Denmark 6.6 11.6 Estonia 9.0 12.9 Finland 6.9 9.3 France 11.2 16.4 Germany 11.5 16.1 Iceland 4.2 6.4 Ireland 11.4 16.1 Luxembourg 13.6 18.9 Netherlands 6.4 9.1 Norway 5.6 8.2 Poland 13.7 15.4 Slovenia 12.5 17.6 Spain 15.0 19.7 Sweden 8.3 9.0 United Kingdom 9.2 16.7

Bulgaria 41.6 50.9

Cyprus 30.2 43.6

Italy 29.8 42.7

Romania 23.4 28.5

(16)

Table 5: The results from the prediction for E. coli and third generation cephalosporins. These are the countries that will reach 20% in the prediction interval. The year when the countries are predicted to reach 20% is listed also with their resistance level in 2016 and their predicted resistance in 2021.

Country Year Resistance in 2016 (%) Resistance in 2021 (%)

Czech Republic 2020 15.1 21.3

Greece 2017 17.6 25.2

Hungary 2017 16.7 27.5

Portugal 2021 16.1 20.3

5.2.4 Klebsiella pneumoniae and Aminoglycosides

Figure 4: Maps of Europe for 2010, 2015, 2020 and 2025 showing the resistance for K. pneumoniae against aminoglycosides in different countries. The resistance level in the Nordic countries are good but for Southern Europe the resistance is already high and will increase.

(17)

In general the uncertainty for this combination is high and only a few countries, including Sweden, had a clear increase. The others were stationary or very varying over the data making it hard to see a trend. Most countries have data from around 2005. Nine countries will not pass 20% in the prediction and seven countries have already passed 20%. Spain might pass 20% in 2019. For Estonia and Slovenia the uncertainty is too high to say where they will be in relation to the 20% limit. Table 6 shows which countries that have not reached 20% as well as those that have and in Table 7 Spain is shown.

In Figure 4 the maps for this combination are shown. We can see that the situation in Southern Europe is much worse than in the Nordic countries, many with a resistance already above 20% in 2015. We can also see that Germany and the UK have a decrease in resistance.

Sweden has an increase unlike the rest of the countries that have not passed 20%, but the resistance is still very low and Sweden will not pass 20% in the next few years.

Estonia passed 20% in 2010 but then dropped below for a few years which makes the model uncertain. The resistance is around 20% right now and might pass above of fall below again but it is hard to say. Slovenia was above 25% in 2009 but has since dropped and is now around 20%. The prediction does not say much. Spain has had a steady increase and according to the prediction will pass 20% between 2017 and 2021, the point forecast passes in 2019.

Table 6: The results from the prediction for K. pneumoniae and aminoglycosides. The first part of the table contains the countries that have not yet reached 20% and the second part the countries that have. For all countries the resistance level in 2016 and the predicted resistance in 2021 is shown.

Austria 4.8 5.7 Denmark 3.2 3.2 Finland 2.3 2.1 Germany 7.8 5.9 Ireland 11.5 16.7 Netherlands 6.1 6.5 Norway 3.3 4.8 Sweden 3.4 4.8 United Kingdom 6.7 6.4

Croatia 36.1 49.2 Czech Republic 47.1 59.9 France 26.2 40.8 Greece 52.9 56.9 Hungary 34.7 45.7 Italy 36.1 48.9 Portugal 35.0 48.4

Table 7: The results from the prediction for K. pneumoniae and aminoglycosides. There is only one country that will reach 20% in the prediction interval, namely Spain. The year when the countries are predicted to reach 20% is listed also with their resistance level in 2016 and their predicted resistance in 2021.

(18)

5.2.5 Klebsiella pneumoniae and Fluoroquinolones

Figure 5: Maps of Europe for 2010, 2015, 2020 and 2025 showing the resistance for K. pneumoniae against fluoroquinolones in different countries. The resistance is higher in Southern Europe and is not predicted to change much in the coming years.

The data stretches from about 2006 and some countries have a steady increase while others, like Denmark and Finland, have stagnated or are decreasing. The level of resistance is also varying, ranging from 5% for Norway (2016) to almost 70% for Greece (2016) see Table 8. Nine countries will not pass 20% and 10 have already passed the line.

In Figure 5 we see that the resistance is high in Southern Europe and for some countries, like Spain, the resistance is increasing. In Sweden the resistance is predicted to decrease from 2015 to 2020.

(19)

Table 8: The results from the prediction for K. pneumoniae and fluoroquinolones. The first part of the table contains the countries that have not yet reached 20% and the second part the countries that have. For all countries the resistance level in 2016 and the predicted resistance in 2021 is shown.

Croatia 43.4 60.6 Czech Republic 50.5 53.6 Estonia 29.5 46.5 France 27.7 44.1 Greece 68.6 80.0 Hungary 35.2 52.5 Italy 56.0 82.6 Portugal 41.7 54.5 Slovenia 29.6 38.5 Spain 22.7 27.7

(20)

5.2.6 Klebsiella pneumoniae and Third Generation Cephalosporins

Figure 6: Maps of Europe for 2010, 2015, 2020 and 2025 showing the resistance for K. pneumoniae against third generation cephalosporins in different countries. The resistance increases slowly and is the highest in Southern Europe.

The data is from 2005 and most countries have an increasing trend, but a few have more varying resistance level between years. Nine countries will not reach 20% in the coming few years. Nine countries have already passed 20%. In Table 9 you can see which countries that have not reached 20% and those that have. In Figure 6, the maps illustrating the resistance for four years are shown. The resistance is higher in Southern Europe but increasing in all countries. The white colored countries are not included, which we can compare with the results from E. coli and third generation cephalosporins where all of the Baltic states and Poland were included.

(21)

Norway and Sweden will not pass 20% in the prediction interval. The predictions for Austria, Germany and Ireland are uncertain since the confidence interval spans from about 5% to 20% for Austria, and for Germany and Ireland it spans over the 20% limit. It is not likely that these countries will pass 20% before 2021 but the uncertainty is large. For Denmark, Finland and the Netherlands the uncertainty is also large but the resistance levels are still well below 20%. The resistance in the UK seems to decrease or at least be stationary, the uncertainty is big but the UK will probably not reach 20% since not even the 95% confidence interval reaches 20%.

For Slovenia the uncertainty is large. It might drop below 20% again, but since Slovenia have been above 20% since 2006 it will probably stay above 20% in the future.

Table 9: The results from the prediction for K. pneumoniae and third generation cephalosporins. The first part of the table contains the countries that have not yet reached 20% and the second part the countries that have. For all countries the resistance level in 2016 and the predicted resistance in 2021 is shown.

Croatia 48.6 54.2 Czech Republic 51.8 64.4 France 28.9 45.9 Greece 72.5 80.9 Hungary 37.5 49.8 Italy 55.8 75.9 Portugal 46.7 59.8 Slovenia 22.8 27.1 Spain 22.3 28.3

(22)

5.3 UTI and RTI Relevant Combinations in Europe

The bacteria P. aeruginosa can cause both UTI’s and RTI’s. Therefore it is presented here.

5.3.1 Pseudomonas aeruginosa and Aminoglycosides

Figure 7: Maps of Europe for 2010, 2015, 2020 and 2025 showing the resistance for P. aerugionosa against aminoglycosides in different countries.

All countries have a decreasing trend or are stationary and almost all are below 20%. The uncertainty in the model is big for this data since there is a large variation, meaning periods of increase and decrease, in the resistance from year to year. Most countries have data going back to 2005. Twelve countries will not reach 20% and two countries are above 20%. In table 10 the countries that have not yet passed 20% as well as those that have passed are presented. For three countries the resistance have been above 20% and just in the last few years dropped below. The countries are presented in Table 11. It is too uncertain to say if Spain will rise above 20% or stay below. Since all other countries show a decreasing trend Spain will likely not rise above 20%.

(23)

In Figure 7 we see the resistance maps for Pseudomonas aeruginosa and aminoglycosides in Europe. In 2010 the resistance is high for some countries and to 2015 it decreases. This decrease in resistance level is predicted to continue.

Spain had an increase between 2005 and 2009 but has since been stationary and slowly started to decline. The decline might continue or the resistance might rise again, but it is too uncertain to say anything. The prediction for Croatia and Greece is uncertain but there is a chance that the countries might drop below 20% in the prediction interval. Czech Republic, Hungary, and Italy have already dropped below 20%. In Table 11 the years when this happened is presented.

Table 10: The results from the prediction for P. aeruginosa and aminoglycosides. The first part of the table contains the countries that have not yet reached 20% and the second part the countries that have. For all countries the resistance level in 2016 and the predicted resistance in 2021 is shown.

Austria 6.1 7.0 Denmark 1.7 4.0 Finland 2.3 -3.0 France 10.7 11.7 Germany 6.9 3.3 Ireland 10.3 6.9 Netherlands 2.8 1.5 Norway 0.9 1.4 Portugal 11.6 13.4 Slovenia 13.3 6.3 Sweden 0.8 2.3 United Kingdom 3.6 2.9

Croatia 33.5 24.9

Greece 28.0 22.1

Table 11: The results from the prediction for P. aeruginosa and aminoglycosides. The countries have been above 20% and in the years listed below they dropped below the 20% limit. Their resistance level in 2016 and their predicted resistance in 2021 is also listed in the table.

Czech Republic 2016 18.8 13.2

Hungary 2016 17.6 15.7

(24)

5.3.2 Pseudomonas aeruginosa and Fluoroquinolones

Figure 8: Maps of Europe for 2010, 2015, 2020 and 2025 showing the resistance for P. aeruginosa against fluoroquinolones in different countries. The resistance in Southern Europe starts high and decreases over the years.

Most countries have data from 2005 and all have either stagnated or are decreasing. Some countries, like Greece, have a high resistance, about 35% in 2016, and others like Denmark have very low resistance, about 4% in 2016. Nine countries are below 20%, two have or will decrease from above 20% to below 20%, and five are still above 20%. For Portugal and Slovenia the uncertainty is too big to say anything. In Table 12 the countries that are below and above 20% are shown and in Table 13 the countries that have dropped or might drop below 20% are found.

In Figure 8 the resistance in Europe for 2010, 2015, 2020 and 2025 is found. In 2010 the resistance is high in Southern Europe but decreases over the years while the Nordic countries do not change much.

Czech Republic has had a decrease and might drop below 20% in the future. In 2015 France dropped below 20% after being at about 25% since the data collection started. Germany has also dropped below 20%, which it did in 2009 and has since stayed below the threshold. Italy has had a decreasing trend and the point forecast passes 20% in 2019. France and Italy are presented in Table 13.

(25)

The prediction for Portugal is very uncertain since the data varies a lot. In 2014-2016 it had a decrease but before that a slow increase, making it too uncertain to say anything about Portugal. Slovenia has danced around the 20% line and the variance is high, meaning that the prediction is uncertain.

Table 12: The results from the prediction for P. aeruginosa and fluoroquinolones. The first part of the table contains the countries that have not yet reached 20% and the second part the countries that have. For all countries the resistance level in 2016 and the predicted resistance in 2021 is shown.

Croatia 37.5 30.4

Czech Republic 34.7 22.0

Greece 34.6 32.2

Hungary 24.3 23.1

Spain 23.0 26.9

Table 13: The results from the prediction for P. aeruginosa and fluoroquinolones. The countries have been above 20% and in the years listed below they dropped/will drop below the 20% limit. The year when the countries passed below 20% is listed with their resistance level in 2016 and their predicted resistance in 2021.

France 2015 13.6 14.6

(26)

5.4 RTI Relevant Combinations for Europe

5.4.1 Streptococcus pneumoniae and Penicillins

Figure 9: Maps of Europe for 2010, 2015, 2020 and 2025 showing the resistance for S. pneumoniae against penicillins in different countries. Sweden is the only country that show an increase. The others are at a very low resistance.

All countries have data from 2005 and many show a decrease. Many are already at a very low resistance and none have really high values. It is only Croatia and Spain that are around 20% and for them the prediction is uncertain.

In Figure 9 the resistance in Europe is shown. The resistance is mostly very low and Sweden is the only country that has an increase. The predictions for Spain and Croatia are uncertain but in 2010 and 2015 their resistance levels were, as mentioned before, above 20%.

Finland is at about zero so the prediction says that it will get negative resistance which is not possible. This is a flaw in our prediction model which we will discuss later.

Croatia has gone from above 20% to zero and up again making the prediction uncertain. It might already have passed 20% since it has varied so much. Spain has the same problem as Croatia, see Table 14.

(27)

Table 14: The results from the prediction for S. pneumoniae and penicillins. The table contains only the countries that have not yet reached 20% because the prediction for the countries that have reached 20% are uncertain. For all countries the resistance level in 2016 and the predicted resistance in 2021 is shown.

Austria 1.1 2.0 Belgium 0.4 -1.6 Czech Republic 0.8 1.4 Denmark 0.4 0.2 Estonia 0 2.4 Finland 0 -0.8 France 0.1 -3.8 Germany 0.7 1.9 Hungary 8.6 4.4 Ireland 0 0.8 Italy 1.3 5.1 Netherlands 0.4 0.3 Norway 1.2 1.5 Portugal 4.2 8.4 Slovenia 0.4 -1.7 Sweden 0.3 9.8 United Kingdom 1.1 0.1

5.4.2 Acinetobacter spp. and Aminoglycosides or Fluoroquinolones

For both of these combinations, the data only stretches five years back in time which is far too little to make predictions from. In the few years of data the resistance varies a lot between countries from around 80% in Italy to around 3% in Denmark for Acinetobacter spp. and aminoglycosides. The same resistance variance and levels are seen for fluoroquinolones as well.

5.5 UTI Relevant Combinations for the USA

The following text summarizes the situation in the USA. The result for the USA is presented separately from Europe since the USA does not have data for the same time interval as Europe. The best time interval for the USA is 1999-2014 making the predictions for the USA stretch only to 2019.

In Table 15 the resistance levels for the last recorded data as well as the last predicted value are shown for all combinations of antibiotics and bacteria.

Table 15: The last recorded resistance value and the last predicted resistance value for different combination in the USA. Three combinations have a prediction to 2019 and three have a prediction to 2017.

Bacteria Antibiotics 2012 (%) 2014 (%) 2017 (%) 2019 (%)

E. coli Fluoroquinolones - 29 - 40.2

E. coli Third generation cephalosporins - 12 - 14.3

K. pneumoniae Third generation cephalosporins - 17 - 20.4

K. pneumoniae Aminoglycosides 11 - 18.4

-K. pneumoniae Fluoroquinolones 14 - 20.9

(28)

-5.5.1 Enterococcus faecalis and Aminopenicillins

The resistance has been at 1% from 1999 to 2012 with the exception of 2009 when the resistance reached 2%. The prediction reaches to 2017 and it will definitely not reach 20% any time soon.

5.5.2 Escherichia coli and Fluoroquinolones

The data is from year 1999 to 2014 and the prediction to 2019. In 2005 the USA passed 20% and for the whole data it has had a steady increase, but has stagnated a bit in the last 6 years. The USA will most likely stay above 20% in the near future.

5.5.3 Escherichia coli and Third Generation Cephalosporins

The USA has data from 1999 to 2014 making the prediction only to 2019. According to the prediction the resistance will stay below 15% throughout the prediction interval. The resistance of USA shows a steady increase which will in the near future probably pass 20% even though our prediction is too short to forecast it.

5.5.4 Klebsiella pneumoniae and Aminoglycosides

The data stretches from 1999 to 2012 and the prediction interval is therefore from 2013 to 2017. There has been some variance in the resistance and from 1999 to 2010 it showed a steady increase, but in the last two years it has had a decrease. The confidence interval passes the 20% line but not the point forecast. Therefore it is a possibility that it will pass 20% in the near future.

5.5.5 Klebsiella pneumoniae and Fluoroquinolones

The data is from 1999 to 2012 and the prediction to 2017. The USA has had an increase and for the last three years it has had a decrease. The point forecast passes 20% in 2016. This means that the USA will probably reach the 20% limit in the near future.

5.5.6 Klebsiella pneumoniae and Third Generation Cephalosporins

In the USA the data stretches from 1999 to 2014 making the prediction until 2019. There has been a large variance in the resistance from year to year making the prediction uncertain. The resistance trend shows a steady increase and according to the prediction the USA will probably reach 20% in the prediction interval and the point forecast passes 20% in 2018.

5.6 UTI and RTI Relevant Combinations in the USA

Table 16: The last recorded resistance value and the last predicted resistance value for the P. aeruginosa combinations in the USA. Both combinations have a prediction to 2019.

Bacteria Antibiotics 2014 (%) 2019 (%)

P. aeruginosa Aminoglycosides 17 15.3

P. aeruginosa Fluoroquinolones 27 27.7

5.6.1 Pseudomonas aeruginosa and Aminoglycosides

The data is from 1999 to 2014 and the prediction from 2015 to 2019 see Table 16. From 2000 to 2006 the resistance was above 20%, but has since 2007 been below. The variance from year to year is high, therefore the prediction is uncertain.

(29)

5.6.2 Pseudomonas aeruginosa and Fluoroquinolones

For all the data from 1999 to 2014 the USA has been above 20% at around 25-30%. It shows a slow decrease but will not pass 20% anytime in the near future.

5.7 RTI Relevant Combinations for the USA

Table 17: The combinations relevant to RTI’s for the USA and their last recorded resistance level and their last predicted value. All three combinations have a prediction that reaches only to 2017.

Bacteria Antibiotics 2012 (%) 2017 (%)

S. pneumoniae Penicillins 17 -0.6

Acinetobacter spp. Aminoglycosides 34 40.6 Acinetobacter spp. Fluoroquinolones 42 57.0

5.7.1 Streptococcus pneumoniae and Penicillins

The data stretches from 1999 to 2012 and the prediction to 2017. In 2008 the USA dropped below 20% and has since stayed there. It might continue to be below 20% or might rise again, but it is too uncertain to say. We can see in Table 17 that the model predicts a very steep decrease because of the variance in the data.

5.7.2 Acinetobacter spp. and Aminoglycosides

The data stretches from 1999 to 2012 making the prediction only to 2017. The USA has been above 20% for all years with data. For the last few years, there has been a decrease from 45% to 35%. There is big variance in the data, making the prediction uncertain. It will not drop below 20%, and even if it did, this is not an antibiotic that should be used as first line since the resistance has been so high.

5.7.3 Acinetobacter spp. and Fluoroquinolones

The data is from 1999 to 2012 and the prediction to 2017. The resistance has varied from about 60% to around 40% making the prediction unclear. We cannot say exactly at what level the resistance will be but it will definitely not drop below 20% in the prediction and this antibiotic should not be considered as first line since the resistance has been so high for a long time.

5.8 A More Detailed Analysis of Four Countries

We did a more in depth analysis of Sweden, France, Italy and the Czech Republic, see Figure 10. These four countries represent different geographical parts of Europe and the resistance level development for E. coli which causes the majority of UTI’s, and S. pneumoniae which is the only bacteria that causes RTI’s that there is data for in Europe. The graphs show the development and we can see that for Sweden the resistance is below 20% and slowly rising for all combinations.

France has a more varied resistance trend for the three combinations. For the E. coli combination the trend is slowly increasing. For E. coli and fluoroquinolones the resistance is predicted to pass 20% in 2017. For the RTI combination, S. pneumoniae and penicillin, the resistance is practically at zero in the prediction.

(30)

Italy is an example of a country with a high resistance level. For both the E. coli combinations the resistance is above 20% long before the prediction even begins. The Czech Republic was chosen because they have many tested isolates each year which is rare for the eastern part of Europe. The resistance in the Czech Republic is higher than in France and lower than in Italy. For E. coli and fluoroquinolones the resistance exceeds 20% in the prediction.

If we look at one combination at a time, we see that the prediction for E. coli and fluoroquinolones is better than the prediction for S. pneumoniae against penicillins. We can see this by looking at the confidence interval, which is narrow. The countries have a clear increasing trend, which is not good for us as humans but it gives a nice prediction. The resistance for S. pneumoniae against penicillins is very low for all coun-tries showing no direct trend. The combination E. coli and third generation cephalosporins have a good prediction for all countries and there is an increasing trend.

We cannot draw any conclusions for RTI’s since we only have one relevant combination. For UTI, the resistance is increasing and since E. coli causes the most UTI’s, there seems to be a need for new antibiotics. But all in all, the resistance trends look different across countries in Europe. Some countries, like Sweden, are still in the safe zone but others, like Italy, are way beyond a safe resistance level. It all depends on where you look.

Figure 10: Four graphs showing the resistance for the E. coli combinations and for S. pneumoniae in four countries. The countries are Sweden, Italy, France and Czech Republic which represents different parts of Europe. The resistance level varies between countries and combinations, shown by the dispersion between countries.

(31)

5.9 Fitness of Model

In order to estimate how well the model fitted our data we calculated the mean absolute error, MAE, for one to five years predictions. This is a good measurement for choosing a good prediction model (Hyndman and Athanasopoulos 2018).

Cross validation is based on splitting up your data into a training set and a test set. The error of the test set is compared with the error of the training data set. This is to see if the model is a good prediction model for this data and if the model is overfitted. If a model i overfitted the error for the training data set is much lower than the error for the test data set, meaning that is is a bad model. There is a time series cross validation function built into the Forecasting package in R. This function did not work for us and therefore we made a script in R that does the same thing. How the MAE-values should be interpreted is hard to know. For most predictions the error increases with the number of years predicted, as it should, and for a first year prediction all but a handful (10) have an error less than 10%. Our interpretation is that under the circumstances, the model suits the data well. But a deeper analysis of our model and the MAE-values is needed in order to evaluate how good it is.

5.10 Number of Tested Isolates

Generally in statistics, the more samples you have the more statistically significant is your model. Therefore, an evaluation on the number of tested isolates was made. In the following graphs all countries are included, even those with less than 50 tested isolates. The graphs show the diversity in number of tested isolates and each line represents a country. You can see that there are many countries which have a very low amount of isolates, but the general trend is that the amount is increasing. An interesting observation is that the USA has, in all graphs, a decrease in the number of tested isolates in the last years. For some combinations the amount increases again but the documentation of data had a dip, why we do not know.

In Figure 11a we see the number of tested isolates for E. coli and third generation cephalosporins. The USA (see blue arrow) starts in 1999 and is the highest in 2000. For 2014 the USA had tested 31,593 isolates and are therefore outside the graph. Germany and the UK (see green arrow) also end up outside of the graph at 15,777 and 21,846 isolates each. France (see red arrow) is steadily increases.

In Figure 11f the number of tested isolates for E. coli and fluoroquinolones are shown. The USA (see blue arrow) starts just below 8,000 isolates in 1999. France (see red arrow), is steadily increasing and had in 2002 tested 2,491 isolates and in 2016 11,251 isolates. Three countries exceeds the boundaries of the plot. The USA which in 2014 had tested 33,879 isolates, Germany (see green arrow) which in 2016 had tested 15,785 isolates, and the UK (see green arrow) which in 2016 tested 22,883 isolates.

In Figure 11h the number of tested isolates for P. aeruginosa and fluoroquinolones are shown. The USA (see blue arrow) comes into the graph between 2005 and 2010. They have a big number of tested isolates starting at 2,318 in 1999 and ending with 11,892 in 2014 with a dip that is seen in the graph. The country with the second highest number of tested isolates is France (see red arrow) and the country that has a high increase in number of isolates is the UK (see green arrow).

In Figure 11k we see the number of tested isolates for Acinetobacter spp. and fluoroquinolones. The number of isolates tested are few and there is only about four to five years of data. There was no prediction made for this combination, except for the USA (see blue arrow). This is because there were not enough data points to make a prediction.

(32)

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j) (k)

Figure 11: The number of tested isolates for each country for the different combinations of antibiotics and bacteria. Each line represents a country. The blue arrows points at the line for the USA and the red arrows points at the line that is France. The green arrow in Figure 11a and 11f points at Germany and the UK. The green arrow in Figure 11h points at the UK.

(33)

6 Discussion

We have a responsibility in how our results are interpreted and therefore we have done our best to discuss all aspects of the results. This includes the difficulties concerning the model and the collection of data. The availability of data was scarce, which limited the length of the prediction interval. The uncertainty of the results is a result of the uncertainty in the data. We will now go into more detail of some of the factors relevant to our project.

6.1 Data Collection

6.1.1 Availability of Data

There is little data on antibiotic resistance levels in general. Most of the data found was presented in pub-lications or articles, but the data was rarely found in tables (i.e. .xlsx or .csv format). This was necessary since we otherwise would have had to transfer the data manually. We made the decision to manually transfer the data from RSN to a .xlsx format. This would not have been possible to do for more sources since it is both time consuming as well as there is a risk of human error. We also reached out to several people working in organizations or projects related to antibiotic resistance, where the main reply was that they did not have data to share and that data is hard to gather since there has not been a common organ collecting data globally. Two factors need to be considered and analyzed when interpreting the result from the prediction model: the number of tested isolates and the number of years containing data. These factors are of importance since the predictive model is based on the previous values and the quality of those will affect the quality of the result. If the number of tested isolates is too low, our model would not have statistical validity.

The data collected from ECDC and CDDEP differs in the number of tested isolates, both between years and countries (see Figure 11). As an example of how the number of tested isolates can differ, one can look at Poland and Sweden. In 2013, Poland had 1,036 tested isolates for E. coli derived infections treated with the antibiotic group third generation cephalosporins. In contrast Sweden, with fewer inhabitants, for the same combination, had 7,532 tested isolates. The prediction for Poland is thus more uncertain than the one for Sweden. Countries also differ in how many years they have collected data on antibiotic resistance.

6.1.2 Invasive Isolates

ECDC and CDDEP have based their data on invasive isolates. The uncertainty with using invasive isolates is that isolates of the same bacteria are not necessarily from the same infections. This means that the antibiotic resistance level for a certain bacteria might not be accurate for different infections. When looking at the result this needs to be taken into consideration. Only looking at samples taken from patients with a UTI or RTI would have been preferable for our model.

6.2 Predictive Model

6.2.1 Possible Improvements

The prediction model proved more difficult to make than we first believed. The field of forecasting is large in economics. There, the data is usually very certain and stretches far back in time and automated functions can be used. This field was new to us, and in the time frame of this course, we were not able to learn everything necessary to develop a model from scratch that would take all aspects into account.

(34)

The model used was a linear model. It does not take into consideration that the values are in percentages, and therefore it predicts negative values and could predict values above 100%. This is not realistic when the values predicted are percentages. Preferably the model would not be able to give those values.

Another possible improvement in the prediction is that, in our model, all years have an equal weight in the model. Ideally, the last few years should have a bigger impact on the predicted value since the resistance for one year depends more on the previous year, than for a resistance level from ten years before. We would also have liked to incorporate the number of tested isolates, where the years with a higher number, weighed more than the years with a lower number.

As mentioned earlier, antibiotic consumption is an important factor in the development of antibiotic resis-tance. So why did we not incorporate this factor in our prediction model? The main reason for it was that although a correlation has been observed, the extent of it has not been mapped and would therefore be difficult to incorporate. This also applies to antibiotic consumption in livestock, where neither the extent of the consumption nor the correlation between usage in animals and the corresponding resistance has been mapped enough (EFSA 2015).

6.2.2 Alice’s Croquet Theory

As has been stated by Dubourg et al. (2017), predictions automatically loses reliability, and sometimes their entire value, when involving biological systems. Parallels can be drawn between this and the Alice’s croquet theory described by Raoult (2016). The Alice’s croquet theory is taken from the book ‘Alice in Wonderland’ and is often used to describe these kinds of predictions. In the book, Alice plays a game of croquet, but instead of having a normal ball and a mallet, she plays the game with flamingos as mallets and hedgehogs as balls. This makes the game incredibly hard to predict, because the game pieces themselves have a will of their own. When predicting future antibiotic resistance the whole human population and several types of bacteria are involved in the prediction. This is why it is important to be aware that there are unknown factors that have not, and possibly could not have, been perceived. All subsequent analyses of the predictions need to take this into consideration. The systems that are involved in the prediction are just as living as the game pieces in the Alice’s croquet theory.

6.2.3 Reach of Our Prediction

There are split opinions regarding how much data is necessary in order to make a forecast and how far that forecast can predict accurately. Rob J. Hyndman, a professor in Statistics (Hyndman, Rob J. 2018), specializing in forecasting, explains how one can evaluate how far a forecast can be based on the number of data points.

“The size of the test set is typically about 20% of the total sample, although this value depends on how long the sample is and how far ahead you want to forecast. The test set should ideally be at least as large as the maximum forecast horizon required.”

- Hyndman and Athanasopoulos (2018)

The total number of samples for each combination spans either 11 or 16 years, with one value per year. Our test set should according to Rob J. Hyndman then be 2.2 or 3.2 values and our maximum forecast horizon would be 2 or 3 years. Since the last collected data was in 2016 our prediction would only forecast until 2018 or 2019.

Rob J. Hyndman also explains how there is no exact limit or rule determining the length of a forecast but that it depends on the variability of the data. If there is a large variability, then more data points are

(35)

necessary (Hyndman and Kostenko n.d.). Looking at previous data of antibiotic resistance some variability can be seen, but since our last data point was for 2016 we wanted to try to predict further than to 2018 and 2019. Based on these arguments our forecast was set to 5 years.

6.3 Presenting Our Data

When presenting your data, it is important that the results are presented in a clear way that is not misleading or difficult to interpret. When we designed our main ways of visualization, this was taken into consideration. The graphs were a way to represent the predictions with the confidence interval, but the challenge was that since there was a substantial amount of predictions it was difficult to get an overview of the results for each country. The maps however were very effective in providing a good overview of the development of antibiotic resistance, but in this case the risk of the visualization being misleading is increased since it did not show the uncertainties. This is problematic since by looking at the map one might assume that the amounts of samples are consistent across Europe and that the error rate of the prediction for each year is the same throughout; both of these assumptions are false. The same reasoning can be applied to the graphs. A large amount of visual results might also be misleading since it is difficult for someone external to assess each of the graphs and maps in terms of their credibility: the more believable results might be drowned in less believable. During this project, a total number of 202 predictions were made and plotted. Because of this, we needed to pick out the ones we assessed to be the most relevant. However, this also poses a challenge: how do we pick out the most important ones and is this selection reliable? The process of picking out the most significant result was based on two things: general representation and quality of data. In this case, general representation means a somewhat varied geographical spread to get an overview. By quality of data, we mean that the country had a decent amount of samples over a decent amount of years. These criteria are however vague and difficult to properly asses, but they give an indication of which countries that are interesting to look at.

7 Conclusion

Based on our prediction, antibiotic resistance is generally increasing in Europe and the USA, both for UTI and RTI related bacteria. It can also be observed by looking at the amount of available data, that this issue is getting more and more prioritized. If we look at the number of tested isolates for each country, they are all increasing year by year. The implementation of GLASS from WHO is an initiative that will further facilitate this.

Our model will hopefully provide Q-linea with the background necessary to make an informed decision in what direction they could develop their products.

(36)

8 Ethical Analysis

The threat of increasing antibiotic resistance is immediate and dangerous. Are we handling the problem in the best way? Are we considering all the ethical implications of the problem? This ethical analysis will consider how to communicate research on this very complex topic, the responsibilities of different parties and a fairness perspective.

8.1 Communicating Research

A question most researchers have to consider is: how do we communicate our results and data to our peers and the public in a clear and correct way? Due to our topic of antibiotic resistance and the threat it poses, this issue is crucial to consider. Even though we are not researchers per se, this is still a scientific project and we needed to consider that when we presented our result. Important aspects to consider as researchers are honesty, objectivity, and openness (Resnik 2013), we highlighted the limitations of our model to decrease the risk of bias. But considering the importance of combating antibiotic resistance in relation to the lack of systematic and comprehensive efforts, how do we communicate our data while still mediating the seriousness of the threat?

This poses a new wave of questions. Do we present the facts and hope people do the sensible thing? Or do we play into fear because we, at this point, cannot afford to wait? Can we justify using manipulation for the greater good? To try to answer these questions, we will discuss two big theories in ethics: consequentialism and deontological, or duty-based, ethics.

8.1.1 Theoretical Aspects

According to Cummiskey (2013), consequentialists at a basic level believe that “right actions maximize good consequences.” What is defined as “good consequences” can vary, but in our case we will define good conse-quences as handling the antibiotic resistance problem, both on an individual level and as a society. Examples of this are decreasing prescription of antibiotics to humans and livestock, educating on proper antibiotic us-age, lowering amounts of prescriptions, and taking legislative measures. With this definition, one could argue that scaring people would be justifiable if it meant that the consequences were good.

On the other hand, if we look at duty-based ethics the opposite can be argued. While consequentialism would argue that the ends justifies the means, duty-based theory would argue that “it is sometimes wrong to promote the best outcome. . . focusing upon actions, rules, maxims, and principles” (Hurley 2013). With this perspective, scare tactics would not fully hold as ethical despite a possible good consequence since, as a principle, the action of manipulation and distorting truth is wrong.

8.1.2 Application of Theory

When studying a problem such as antibiotic resistance one needs to consider the responsibility of presenting data and findings. Obviously the results should be presented truthfully, but the truth can be presented very differently. An example with antibiotic resistance is that it is a scary problem. Therefore one could present the result of increasing resistance in a "the end is nigh"-manner with focus on approaching a return to the pre-antibiotics era. A time where an infected wound or illness that is simply treated with a few pills today could cause not only deaths but "your death" and "your family and friends’ deaths". In contrast to this however, the results could be presented as an issue that simply needs to be taken seriously and improved but not as an impending doom. The main difference between these approaches is inciting fear in the general public. As previously mentioned, inciting fear can arguably seem like the ethical approach, but one can argue the opposite.